More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents

More Memory, Worse Agents. Error Reproduction and Anti-Persistence in LLM Agents. By Ruhan Wang and two others, from Indiana University and Tencent HY. Published on the nineteenth of October. No code repository is linked to it.

An LLM agent can get better across tasks without any change to its weights. It stores experience from its own executions in external memory and reuses that experience to guide later decisions. One design question sits at the centre of this. Should memory hold only the observed trajectories and their outcomes, or also the rules, summaries and skills inferred from them? Keeping those abstractions has an obvious appeal: strategies can be reused directly, instead of being derived again from raw trajectories each time.

The cost is the paper's subject. An error inside a stored abstraction does not stay put. It propagates into later decisions through retrieval and reuse, which is the familiar risk. The authors name a second and less obvious route. A faulty abstraction also shapes the execution trajectories the agent produces, and those trajectories then become the raw material for new abstractions, which reproduce the original error. They call this error reproduction: an error regenerated through a cycle of abstraction, execution, and re-abstraction, so that its life extends past any direct reuse of the faulty knowledge it came from.

Anti-Persistence is their memory design, and it cuts the cycle by refusing to store abstractions. Trajectories and outcomes accumulate as before. Guidance specific to a task is reconstructed on demand instead of retained. An adaptive selector reads prior outcomes to decide whether to synthesize guidance at all, and how. Once execution is finished, that guidance is discarded. Nothing is left behind that a later retrieval could pick up and treat as settled knowledge.

What this buys is precise, and the authors are careful about it. Discarding the guidance prevents it from being reused directly as a stored object. Indirect feedback through the retained trajectories is still possible, so the loop is narrowed rather than sealed.

Across different benchmarks and different backbones, Anti-Persistence reaches higher overall held-out success than the other memory designs evaluated, and it does so at the lowest average cost in LLM calls among them. Reconstructing guidance each time might have been expected to cost more calls, not fewer, so the cost result is doing some work alongside the success rate.

The title's provocation lands on a distinction rather than on memory itself. Accumulating experience and keeping the conclusions drawn from experience are two different things, and the results support separating them in memory-based self-improvement. Let the record persist, and let the inferences be made fresh and then let go.